Research On Indoor Positioning Based On Visual Point And Line Feature And Inertial Fusion

Indoor intelligent robots have the ability to perceive the surrounding environment and perform autonomous navigation and positioning in unknown and complex indoor scenes.Its core technology is to perform real-time and accurate state estimation of the robot.The visual inertial odometer can realize the demand of indoor navigation and positioning only by fusing the sensor information of the camera and the IMU,which has been widely studied.In view of factors such as large areas of weak texture areas,simple repeated texture areas,and long corridor lighting changes in the indoor structured environment,traditional solutions based only on visual point features or visual point features and IMU due to the failure of point feature matching can easily lead to a decrease in positioning accuracy.This paper introduces the rich line features in the indoor environment,improves the point-line feature extraction and matching algorithm,and builds fusion graph optimization nonlinear models based on point-lineIMU information to improve the positioning accuracy and robustness.The main research work of this paper is as follows:(1)The extraction and matching algorithm of visual point and line feature is studied.Violent matching of point features has a large number of mismatching problems,and the traditional KNN algorithm has the problem of fewer effective matching point pairs.This paper studies the GMS algorithm based on the grid motion statistics strategy,which is based on the motion smoothness constraint condition,and statistically compares the similarity of the feature points in the grid to match and remove the feature points.In the comparative analysis of the indoor scene images of the TUM dataset,it is concluded that this method can save more effective matching point pairs while effectively eliminating false matches.For line feature extraction,there is a large amount of calculation and the problem of line segmentation.This paper studies an improved line feature detection and extraction strategy.Before LSD extraction,the pre-processing mechanism of the gradient density filtering mechanism is used to extract line features only for the borders of densely line feature areas instead of all areas to reduce the amount of calculation.Moreover,the angle characteristics between the line segments and the relative spatial position characteristics are used to merge the similar broken lines,which improves the extraction effect of the line segments.(2)A pose estimation method based on point-line-IMU information for tightly coupled graph optimization model is studied.This article introduces the IMU preintegration and the projection measurement model of point and line features,and derives the IMU pre-integration residual model,point feature residual model and line feature residual model.Then,with the help of the factor graph model,a graph optimization model for pose estimation for multi-state quantities is constructed.The Eu Ro C data set is used to verify the positioning feasibility and positioning accuracy of this method,and the parameters such as mean error and root mean square error are compared,which shows that the method has high trajectory positioning accuracy and relatively stable performance.(3)A complete set of indoor positioning experiment platform integrating point and line features and inertial navigation is realized.The overall framework includes four parts: data preprocessing,initialization,sliding window nonlinear optimization mechanism to estimate pose,and closed loop.Compared with the traditional solution,the positioning accuracy on the Eu Ro C dataset is improved by about 10% on average.In the actual scene data,the positioning track error of this solution is about 0.6%,but the error of the unimproved scheme reaches about 2.8%.In the comprehensive comparison,the improved scheme improves the positioning accuracy and robustness.